Analog AI Is Back, But Can It Survive Its Own Noise?

Analog computing is re-emerging as a solution to AI's significant energy consumption by performing matrix multiplications directly within memory using physical laws. This in-memory computing approach avoids the energy cost of moving data, but introduces inherent physical noise that can degrade model accuracy. Simulations show that a conventionally trained network's performance collapses when faced with this noise. The solution is hardware-aware training, which injects simulated noise during the training process itself. This forces the model to develop weights that are robust to the physical imperfections of the analog hardware, maintaining high accuracy despite the noise.